GAL-UNet: lightweight detection of diverse defects in fabric with imbalance resilience

Abstract The textile industry, critical to product quality and economic prosperity, demands sophisticated fabric defect detection systems. This paper introduces GAL-UNet, an innovative lightweight model adept at real-time processing, diverse defect identification, and overcoming dataset imbalances. Firstly, GAL-UNet integrates a Group multi-axis Hadamard Product Attention (GHPA) module to streamline the model architecture, significantly reducing parameters and accelerating inference while maintaining high detection accuracy. Second, GAL-UNet is fortified with three advanced attention mechanisms: a Non-Local module for capturing global contextual information, a Parallel Pathway Attention Gates (PPAG) module for spatial focus within skip connections, and a Squeeze-and-Excitation (SE) attention module for channel-wise feature recalibration. Collectively, these mechanisms enhance feature extraction, significantly improving the detection of defects with notable differences. Furthermore, GAL-UNet introduces an optimized loss function that incorporates the pixel count of defects, effectively mitigating sample imbalance and focusing on less prevalent defect types. Experiments demonstrate GAL-UNet’s superior performance, achieving a detection accuracy of 95.58% and surpassing state-of-the-art models like DeepLabv3+ and Attention UNet. The model’s real-time processing capabilities make it well-suited for precision and speed requirements in industrial fabric inspection. Our code and data are publicly available for further research and application development at the website: https://github.com/kingkamin/GAL-UNet .